Model Cards & System Cards

In 2019, a team of researchers proposed a simple discipline: every released model ships with a model card — a standardized sheet reporting intended use and out-of-scope uses, what data trained it, how performance breaks down across demographic groups (not just the aggregate score), and the evaluation setup.

The insight was that a headline accuracy number hides the thing that actually matters: for whom does it fail? A face-recognition model at 99% aggregate accuracy could still be much worse for darker-skinned women — a fact only visible in the disaggregated rows the card forces you to publish.